Popularity bias is a phenomenon associated with collaborative filtering algorithms, in which popular items tend to be recommended over unpopular items. As the appropriate level of item popularity difers depending on individual users, a user-level modification approach can produce diverse recommendations while improving the recommendation accuracy. However, there are two issues with conventional user-level approaches. First, these approaches do not isolate users' preferences from their tendencies toward item popularity clearly. Second, they do not consider temporal item popularity, although item popularity changes dynamically over time in reality. In this paper, we propose a novel approach to counteract the popularity bias, namely, matrix factorization based collaborative filtering incorporating individual users' tendencies toward item popularity. Our model clearly isolates users' preferences from their tendencies toward popularity. In addition, we consider the temporal item popularity and incorporate it into our model. Experimental results using a real-world dataset show that our model improve both accuracy and diversity compared with a baseline algorithm in both static and time-varying models. Moreover, our model outperforms conventional approaches in terms of accuracy with the same diversity level. Furthermore, we show that our proposed model recommends items by capturing users' tendencies toward item popularity: it recommends popular items for the user who likes popular items, while recommending unpopular items for those who don't like popular items.
• Information systems → Personalization; Recommender systems; popularity bias, temporal information, personalized recommendation
Recommender systems help users to access the specific
information that they seek from a huge amount of data. Accurate
recommendations lead to an increase in customers’ purchases or
consumption; hence, there is a need for more eficient recommender
systems that produce personalized content for individual users.
TempRRS ’17, August 2017, Como, Italy
Copyright © 2017 for this paper by its authors. Copying permitted for private and
academic purposes.
To produce personalized recommendations, collaborative filtering
(CF) is a widely used approach. The CF approach produces items
for a target user using data compiled from observations of users
with similar preferences as the target user [
In the CF approach, it has been noted that popular items tend
to be recommended more often [
However, there are two issues in conventional user-level approaches. First, these approaches do not isolate users’ preferences from their popularity tendencies clearly. Second, although item popularity changes dynamically over time in reality, these approaches do not consider temporal item popularity. In general, incorporating temporal item popularity into models improves the recommendation accuracy. Moreover, to counteract popularity bias, especially in user-level solutions, incorporating temporal item popularity is important because the reasons of users’ behaviors are considered diferent depending on their purchase time even if they purchase same items. To the best of our knowledge, however, there is no approach considering temporal item popularity in the field of counteraction against popularity bias.
In this paper, we propose a novel approach to tackle the popularity bias, namely, MF-based CF incorporating item popularity orientation of individual users. Our model isolates users’ preference from their tendencies toward item popularity clearly. We also consider temporal item popularity and incorporate it into our model. To verify the eficacy of the proposed model, we conducted experiments using a real-world dataset. The experimental results show that our model improves both accuracy and diversity compared with a baseline algorithm in both static and time-changing models. Moreover, our model outperforms conventional approaches in terms of accuracy with the same diversity level. We also demonstrate that our proposed model recommends items by capturing users’ tendencies toward item popularity: it recommends popular items to users who like popular items, and unpopular items to those who do not like popular items.
We summarize the main contribution of this paper as follows: Our model isolates users’ preferences from their tendencies toward item popularity clearly.
We consider temporal item popularity in the field of counteraction against popularity bias.
We conduct experiments using a real-world dataset to verify the eficacy of the proposed model. 2 2.1
Popularity bias is a phenomenon of existing recommendation
algorithms in which popular items tend to be recommended over
unpopular items. To tackle this problem, several approaches have
been proposed [
Global-level approaches modify their recommendations for all
users uniformly [
User-level approaches consider these diferences and then
modify their recommendations depending on the individual user’s
tendencies toward item popularity. Therefore, user-level approaches
possibly improve both diversity and accuracy simultaneously. The
conventional user-level approaches proposed in [
Incorporating temporal aspects into CF has been investigated,
particularly for developing accurate recommendation algorithms. For
example, [
In this section, we present our MF-based model that incorporates individual users’ tendencies toward item popularity. We focus on situations where personalized top-N recommendations are produced based on users’ implicit feedback (e.g. views, clicks, purchases, etc.). 3.1
In MF, both items and users are characterized by vectors of latent factors derived from explicit feedback (e.g. ratings) as well as implicit feedback. The basic model of MF with item bias is formulated as follows: xˆui = bi0 + fuT fi ; (1) where xˆui is the prediction score of preference of user u toward item i, bi0 is an item-specific bias which represents item popularity, and fu and fi are k-dimensional vectors of latent factors of user u and item i, respectively. The inner product fuT fi achieves a high value when both user and item vectors are similar. Furthermore, item bias bi increases when an item is popular. The prediction score is determined by their aggregation.
If item bias bi values are extremely high, the item is recommended regardless of whether users like it or not. Hence, recommendation systems tend to recommend these items, which leads to popularity biased recommendation. A simple solution for this problem is to penalize items according to the item popularity. However, preference toward popular or unpopular items varies for each user. Considering this, the solution is not suitable for users who like popular items. Therefore, the penalization of popularity needs to be changed depending on the users’ popularity tendencies.
Moreover, the users’ popularity tendencies should be
considered along with the items’ temporal aspects for two reasons. Firstly,
item popularity changes dynamically over time in the real world
for various reasons [
Therefore, we develop a model to incorporate both users’ popularity tendencies and items’ temporal popularity, which is formulated as follows: xˆui = ¹bi0 + bi ¹t ºº¹1 + дu º + fuT fi ; (2) where дu is the user-specific parameter of popularity tendency and bi ¹t º is the time-varying item bias at the period of time t . The parameters, bi0, bi ¹t º, дu , fu , and fi , are learned by optimization.
The дu value works as the balancing parameter between the item popularity and preference toward the item. When the дu value of a user u is greater than zero, the user prefers popular items to unpopular items. High дu values indicated that the user may simply prefer popular items without regard to his/her item preference. Conversely, when it is less than minus one, the user prefers unpopular items to popular items.
As mentioned in Section 2.1, users’ preferences and their tendencies toward item popularity are mixed in conventional userlevel approaches. In contrast, our model resolves the confusion by modeling as in Eq. 2: the first term represents item popularity and users’ popularity tendencies, and the second term represents item feature and users’ preference. Therefore, our model captures these features separately. 3.2
Our model formulated in Eq. 2 can be learned by applying
existing optimization methods, such as point-wise and pair-wise
optimization. For example, for point-wise optimization, root mean
square error (RMSE), which is used in Biased-MF [
In this section, we conduct experiments using a real-world dataset to verify the eficacy of the proposed model. 4.1
We used the Amazon.com Movies and TVs dataset [
In our experiments, we performed five-fold cross validation and aggregated the results. First, we randomly selected 80% of observed feedback as a training set to train models, and the remaining 20% as the testing set for the trained models. To measure the performance, we used three evaluation metrics: the top-N prediction precision (Precision@N), the top-N prediction recall (Recall@N), and the top-N item coverage (Coverage@N). We set Iupred¹t º as the predicted items of user u over a certain period of time t , and Iutrue t ¹ º as the true list in the testing set. Prediction is performed for each period of time t , and each user’s scores are aggregated over each where jIupred¹t ºj = N , U is the set of users in the testing set and Tu is the set of the period of time when interactions of user u are observed. Similarly, the top-N prediction recall is defined as: 1 ∑ 1
∑ jU j u 2U jTu j t 2Tu jIupred
I true¹t ºj j u ¹t º \ Iutrue¹t ºj : The top-N item coverage applies to all the output that a recommender system produces for a set of users. This metric is also called the top-N aggregate diversity. In our experiment, this metric is deifned as: ∑
∪ t 2T j u 2Ut Ru j ; jT j where Ut is a set of users whose interactions are observed at a period of time t and Ru is the recommendation lists for user u, and the length of the lists is N . 4.3
To examine the performance of our proposed methods, we
compared them with conventional approaches. For the optimization
of our methods and base models of conventional approaches, we
selected the Bayesian personalized ranking (BPR) procedure [
re-ranking method that considers an individual’s personal popularity tendency (PPT). It balances novelty and user preference by matching the PPT of a recommendation to that of the users measured by earth movers distance (EMD), which is a distance metrics between two distributions.
re-ranking method that considers an individual user tendency suitable for multiple optimization goals. In our experiments, the optimization target is set to EMD.
BPRMF(t)-pop is the method proposed by this paper in Eq. 2. BPRMF-pop is the model that removes temporal item popularity from Eq. 2.
To model PPT, the discrete distribution of the binned popularity values of the items is required. In our experiments, we defined the item popularity of the recommendations as the number of item occurrences in the top-N recommendation lists for the active users. We used a log-scaled popularity histogram for discrete distribution. The parameters of all models were tuned so as to maximize the accuracy metrics. In the case of conventional approaches, it is known that a higher coverage setting reduces accuracy. Hence, we selected the parameter value for which the coverage score became close to that of our model.
Methods
Precision@10
Recall@10
Coverage@10 BPRMF +PPTM (c = 0:1) +PRA (Xu =5) BPRMF-pop BPRMF(t) +PPTM (c = 1) +PRA (Xu =5) BPRMF(t)-pop
Table 1 shows the results of the comparison between our method and conventional approaches. The number of latent factors was set to 300 and the number of items in a recommendation list to 10. In general, time-aware models improve accuracy and reduce coverage compared with static models. Our model improved both accuracy and diversity compared with the baseline in both static and time-changing models. Particularly in case of time-varying models, our model achieved significant improvement. This indicates that considering temporal item popularity is essential to capture users’ tendencies. Our model outperformed conventional approaches in terms of accuracy with the same diversity level. Therefore, our model efectively captures users’ preference and their tendencies toward item popularity.
We suppose that the interactions of users with mainstream tastes are easy to predict. As our model isolates users’ preference from their tendencies toward item popularity, we can verify the idea by analyzing the distribution of accuracy depending on the magnitude of дu values. Figure 1 shows the plots of дu values versus two evaluation metrics: Precision@10 and Recall@10; each point is the average of evaluation metrics with regards to the average of дu values of 100 users in descending order of the дu value. As can be seen from Fig. 1, both Precision@10 and Recall@10 of the users who have large дu values are high. Therefore, this result supports our assumption.
We also investigated the relation between users’ purchase behavior, which corresponds to their tendencies toward temporal item popularity, and our model’s recommendations. Table 2 shows
(a) Item popularity orientation score дu = 0:69, a user likes popular items. bi ¹t º (#rank)
Pref. score bi ¹t º (#rank)
Pref. score TopN items TopN items the examples that our model recommends popular items for the user who likes popular items and vice versa. bi ¹t º score represents item popularity at the period of time t and actual users’ purchase is shown in bold in Table 2. As can be seen from the user’s purchase behavior shown in Table 2-(a), the user tends to purchase popular items. Our model learned such purchase behavior from the user’s past purchases, and then evaluated the дu value of the user as 0:69, which means that the user likes popular items. Our model produced popular items for the user, which were ranked in the top 20. On the other hand, the user in Table 2-(b) selected items that match the user’s preference without regard to the items’ popularity. Our model captured the tendency from the user’s past purchases and evaluated the user’s дu value as 1:19. Our model recommended items that match the user’s preference regardless of their popularity for the user. The preference scores are all high, while the items’ rankings are various. Therefore, these results indicate that our model captured the users’ popularity tendencies and recommended personalized items appropriately.
In this paper, we proposed a novel approach for counteracting popularity bias, using MF-based CF incorporating individual users’ tendencies toward temporal item popularity. Our model isolated users’ preference from popularity tendency clearly, and considered temporal item popularity. The experimental results based on a realworld dataset showed the eficacy of our model.
In future work, we plan to further verify the efectiveness of our proposed model by using various datasets in diferent domains or by learning other optimization methods for top-N recommendation. Moreover, as well as item popularity, users’ tendencies toward item popularity may change over time. We plan to investigate this temporal phenomenon.